In briefRankBuffer: Efficient Ranking-Based Rewards for Open-Ended Generation
SummaryThis paper introduces RankBuffer, a method for building rewards in reinforcement learning for open-ended text generation, where no single correct answer exists. It keeps an ordered, per-query buffer of previously judged responses, first placing each new response into an anchor interval through a coarse judgment and then ranking only responses within the same interval. The full order is converted into bounded rank rewards. Across four open-ended benchmarks, RankBuffer outperforms all pointwise baselines and nearly matches the strongest ranking-based baseline while substantially reducing judging cost.
In briefFocusVTC: Efficient, High-Performance Visual Text Compression with Adaptive Resolution
SummaryThis paper proposes FocusVTC, which renders long text as images to shorten inputs to large language models, using adaptive resolution to avoid the trade-off between token savings and legibility. It pairs low-resolution global views with selective enhancement of relevant regions, trained via supervised fine-tuning and reinforcement learning. The authors report gains on RULER, LongBench, MRCR, and VTCBench, a 2.79x online speedup over text input, and preserved multimodal performance on MMMU and MME.
In briefVLM4Cluster: A Benchmark for Deep Image Clustering with Vision-Language Models
SummaryThis paper introduces VLM4Cluster, a comprehensive benchmark for image clustering in the era of pre-trained vision-language models. It implements 17 representative methods and evaluates them on 20 datasets covering classical, challenging, fine-grained, large-scale, and out-of-distribution settings. The study also examines robustness to adversarial perturbations, generalization under distribution shifts, and computational efficiency. Results show that language-assisted image clustering performs strongly on semantically demanding benchmarks and generalizes better under distribution shifts, but its gains are less consistent on large-scale and fine-grained datasets, and language assistance does not systematically reduce adversarial sensitivity.
In briefHow the target of predictive supervision shapes what VLA policies learn
SummaryThis paper studies how predictive supervision affects the visual representation learned by vision-language-action (VLA) policies. In controlled comparisons with matched target constructions, prediction horizons, and training conditions, different prediction interfaces produced markedly different forecasts and visual representations. VLA policies trained with more direct, scene-matched future supervision showed stronger robustness under simulated and physical distribution shifts.
In briefOCA: ODE-driven cross-attention for matching images to point clouds
SummaryThis paper addresses attention ambiguity in learning-based image-to-point-cloud registration by modeling the ideal feature interaction with ordinary differential equations. It proposes an ODE-driven cross-attention (OCA) module that refines feature representations and attention matrices. OCA can be integrated into existing registration frameworks, and tests on five baselines and four public datasets show registration recall gains of up to 5%, 9%, and 15% in standard, fine-tuning, and zero-shot settings.
In briefPR-OPD: On-policy self-distillation with privileged representations for agent reinforcement learning
SummaryThis paper studies reinforcement learning for language-model agents, where a policy given a skill teaches its skill-free version through token probabilities. It identifies two phenomena: a skill raises WebShop success while changing few token probabilities, yet alters hidden states of over 80% of response tokens. The authors propose PR-OPD, which after a GRPO warm start aligns projected hidden states with a hindsight-skill teacher at every layer. On ALFWorld and WebShop with two backbones, it achieves the best overall results, improving ALFWorld success by up to 4.7 points over GRPO.
In briefInducing Process Supervision from Outcome-Only RL (A method for training step-level process reward models using only final-outcome signals, without per-step annotation.)
SummaryProcess reward models give step-level feedback for LLM reasoning, but training them is costly because human step annotation is hard to scale and Monte Carlo estimation is computationally expensive. The paper introduces TIPS, an outcome-only reinforcement learning framework in which the model produces a chain-of-thought, step labels, and an outcome label, with reward based solely on outcome correctness. The authors validate TIPS on math and agent benchmarks across four backbone families. TIPS-Qwen3-4B-Thinking-2507 reaches 85.2 F1 on ProcessBench using only 3.2K outcome-labeled trajectories.
WhyNovel idea of inducing step-level supervision from outcome-only RL; ProcessBench F1 of 85.2 supports it, though evaluation is limited to math and agent benchmarks.
In briefA physics-based digital toolkit that simulates realistic scanning tunneling microscopy images for AI training
SummaryScanning tunneling microscopy (STM) images atomic-scale surfaces but is highly prone to noise. This paper presents a low-cost, physics-driven toolkit that simulates clean STM images and adds physically motivated artifacts and noise, such as Gaussian electronic noise, 1/f flicker noise, scan-line noise, background tilt, and mechanical drift. The authors show the generated data can train denoising models for Cu(111) and Pb(111) surfaces while preserving atoms, defects, and electron waves. Supervised models trained on this data outperformed benchmarked unsupervised methods at denoising and learning electron wave patterns on Cu(111) images.
Huanhuan Zhao, Laxmi Bhurtel, Connor Vernachio et al.
In briefPE-OPSD: Building Prompt Enhancement into Image Generators Through Self-Teaching
SummaryText-to-image models often work best with detailed prompts, but users usually write short ones. Existing prompt enhancers rewrite prompts at inference time, adding latency. This paper proposes PE-OPSD, which trains a raw-prompt student on its own generation path, guided by a teacher that uses enhanced prompts. At inference, the enhancer and teacher are removed, so the model generates directly from raw prompts.
In briefIs Bayes-Sufficient Compression Enough? How Communication Aids Multi-Agent Systems
SummaryThis paper studies when a short message from a sender with broad context improves a bounded executor's next decision in multi-agent LLM systems, compared with passing raw context or using a stronger sender. It proposes a framework called receiver-relative bounded coordination, which treats message utility as receiver gain minus protocol cost. Across six benchmarks, the same protocol raises accuracy on one task while lowering it on another, and the authors add an inference-time selector that improves the accuracy-cost trade-off in the evaluated settings.
In briefWhat makes recurring computation useful in looped language models?
SummaryThis paper studies looped language models, which increase computational depth by reusing shared parameters without adding new ones. Through controlled experiments, the authors examine when recurrence helps, where it should be applied, and how its conditioning affects performance. They find recurrence can improve reasoning beyond the training horizon but can hurt knowledge tasks. They propose history-state injection combined with timestep conditioning as a low-cost design that better preserves knowledge and handles varying inference budgets.
In briefWitnessGym: Benchmarking Coding Agents on Building Executable Bug Evidence
SummaryThe paper presents WitnessGym, a framework that automatically builds bug-validation benchmarks by injecting bugs into test-reached paths of real Java projects and keeping cases exposed by a construction-time witness. It produces 1,300 benchmark cases whose injected patches resemble historical bug patches and are hard for two evaluated models to tell apart in blinded comparisons. Four coding agent frameworks across six framework/model pairings are evaluated, and the authors report that constructing witnesses remains difficult even when the bug pattern is known.
In briefPhysics-aware machine unlearning for cyber-physical systems, removing harmful learned behavior while keeping physical rules intact.
SummaryThis paper proposes a gradient-ascent machine unlearning method that uses the physical residual of a cyber-physical system to steer weight updates toward physically feasible regions. It is tested on an IEEE 34-bus distribution system with two physics-informed neural network controllers, validated by OpenDSS power-flow co-simulation. The method is compared against naive gradient ascent, exact unlearning, SISA, and full retraining. The authors report that it removes poisoned behavior while restoring physical compliance.
Mohammad Zakaria Haider, Muhammad Nadeem, and Mohammad Ashiqur Rahman
In briefAgent Distillation Roadmap: Transferring Knowledge Across Models, Artifacts, and Harnesses
SummaryModern agents rely increasingly on memory, tools, and execution logic, not just model parameters. The paper defines Agent Distillation as the persistent transfer of task-solving knowledge from a teacher agent to a student agent. It organizes the field by where transferred knowledge is retained: within the model, as artifacts, through the execution harness, or across substrates. It also proposes an evaluation framework linking retention to causal contribution and deployed utility.
In briefGraded Preference Optimization for Accurate Limb-Motion Video Captions
SummaryVision-language models often assign actions to the wrong person or limb in video captions, especially across camera cuts. The authors introduce FlexBench, a benchmark with 3,105 shots and 18,161 evaluation queries, and a Graded Physical Alignment score that rewards correct actions and penalizes incorrect or fabricated ones. They also propose GM-DPO, a training method that gives larger preference margins and weights to more severe action errors. Across three backbones, GM-DPO improves GPA over DPO by 2.02 to 3.40 points.
In briefSemantic projection method for continual self-evolution of language agents
SummaryLanguage-model agents rely on persistent natural-language skills to adapt, but repeated revisions to a shared skill can overwrite procedures needed for earlier tasks. The paper introduces Semantic-Scope Projected Evolution (SSPE), which transfers gradient projection from parameter space to behavior space. SSPE identifies acquired capabilities that a proposed skill revision may interfere with, and uses observed gains and regressions to build a compatible revision instead of simply rejecting it. Across controlled synthetic streams and heterogeneous real-agent benchmarks, SSPE improves final cross-domain competence and reduces forgetting relative to strong baselines, and retains the best average performance after transfer to a different executor model.
In briefDualTrack: Synchronized Speech and Gesture Generation via Coupled Pretrained Models
SummaryThis paper presents DualTrack, a method that couples pretrained speech and motion models on a shared 12.5 Hz timeline to generate speech and gestures together. Evaluated on 43 BEAT2 recordings in four languages, it reports on shared English and Spanish inputs lower word error rate and full-motion Fréchet Gesture Distance, and higher beat consistency and speech naturalness, than the evaluated GELINA baseline.
In briefNeural Structural Reasoner: a brain-inspired model for reasoning over structured knowledge
SummaryThis paper introduces the Neural Structural Reasoner (NSR), a brain-inspired network that keeps relational structure in the connectivity and dynamics of coupled neuron populations, and uses it for link prediction on knowledge graphs. On standard knowledge-graph benchmarks, NSR achieves competitive accuracy but does not lead on every dataset, and its reported training times are lower than several neural baselines. Because reasoning is carried out through human-readable sequences of neuron activations, intermediate inference steps can be tracked, and the model can extract latent relational hierarchies and compositional rules.
In briefSemPSG: A Semantic Channel-Aware Foundation Model for Polysomnography Analysis
SummaryPolysomnography records multiple physiological signals to characterize sleep, but channel configurations differ across centers, making transferable representation learning difficult. SemPSG explicitly encodes the physiological meaning of each channel, based on signal modality and reference setup, and uses it in both signal representation and channel aggregation. It pairs a semantic-conditioned time-series encoder with a multi-view image encoder. The model is evaluated on sleep staging, sleep-disorder breathing analysis, disease prediction, cognition and emotion recognition, and demographic estimation, where it outperforms general time-series and PSG-specific foundation models and generalizes across datasets.
In briefGenerating questions that prompt revisions for AI research papers
SummaryThis paper studies whether large language models can generate edit-inducing questions, meaning questions whose answers would improve a paper draft. Using paired submission and camera-ready papers from ICLR and NeurIPS, the authors compare GPT questions with and without full paper context against those from human reviewers. GPT produces more questions tied to more extensive and broader edits, but a much smaller share of its questions are edit-inducing.
Sebastian Joseph, Zichao Wang, Jennifer Healey et al.
In briefCrossTimeEdit: A Decade-Spanning Dataset and Reward-Guided Model for Editing Historical Street Views
SummaryHistorical street-view imagery has gaps, so the authors built VIGOR-his, a dataset of 43,653 location-level quadruplets across 11 cities on three continents. They frame generating past street views as an editing task, using recent street views and satellite change evidence as constraints. The CrossTimeEdit model, built on FLUX.2 [Klein] 4B, is trained with supervised fine-tuning followed by reinforcement learning using three criteria. The paper reports a 17.12% overall improvement over the pretrained baseline.
In briefCI-PINN: A causal integral neural network for solving evolution equations
SummaryPhysics-informed neural networks (PINNs) solve partial differential equations by adding physical laws to the training loss, but their standard pointwise space-time form does not explicitly capture temporal dependence for evolution equations. This paper proposes CinNet, a neural architecture whose core is a Volterra-type causal integral term that aggregates historical features to encode temporal dependence. Based on CinNet, the authors build CI-PINN, which in benchmark experiments achieves higher solution accuracy than various PINN variants, especially with sparse collocation points.
In briefSelective Questioning as a Hidden Commercial Influence on Shopping AI
SummaryThe paper studies whether commercial incentives can steer a shopping assistant by changing which preference question it asks, rather than its final ranking. In a small synthetic setup, an honest simulated user answers one pairwise question, and a recommender sees only the products and the answer, not the sponsorship. A targeted adversarial instruction raised sponsored selections by 0.30 and lowered synthetic utility by 0.0547 versus neutral questioning, while a soft commercial instruction changed no selections. The authors state the results do not show typical behavior under advertising incentives or effects on real consumers.
In briefDo LLMs Really Forget? Hidden-State Leakage in Unlearning and Its Fix
SummaryThis paper argues that unlearning in large language models is usually judged only by outputs, so a model can look like it has forgotten sensitive content while that information remains in its hidden representations. The authors show this with probe decoders trained on hidden states across layers, recovering substantial sensitive information on the TOFU, MUSE, and WMDP benchmarks even when output metrics indicate success. They propose Probe-Adversarial Representation Suppression (PARS), which adversarially minimizes extractable information from hidden representations. According to the paper, PARS gives stronger erasure guarantees under probing and relearning attacks than all evaluated baselines.
WhySeparates output suppression from representational erasure and proposes adversarial probing-based suppression, validated across three benchmarks with layer-wise probes.
In briefReducing source-reliability shortcuts in RL fact-checkers
SummaryRetrieval-augmented fact-checkers often receive HIGH or LOW trust labels for evidence sources, which should affect confidence and search but not the verdict. The paper introduces TrustSwap, a counterfactual test that swaps, lowers, or removes these labels while holding evidence text fixed, measuring verdict, confidence, and search separately. It finds that label changes alone flip some confident verdicts, and standard GRPO fine-tuning amplifies this at 8B. The proposed trust-swap augmentation (TSA) reduces verdict flips at 4B, with no detectable effect at 8B.
In briefFaster, Compressed Agnostic Federated Learning to Cut Communication Costs
SummaryThis paper studies agnostic federated learning, which aims for a model that performs reliably across many heterogeneous workers while limiting costly communication. It proposes AFL-BR, which uses online mirror ascent with KL divergence and blockwise restarts to reduce synchronization rounds, and AFL-Com, which adds bidirectional compression with error feedback to lower per-round communication. The authors give convergence and communication complexity results and report experiments validating the improved efficiency.
In briefA practical guide to EEG analysis: preprocessing, modeling, and machine learning best practices
SummaryThis chapter is a practical methodological guide to EEG analysis, covering preprocessing steps such as artifact removal, filtering, bad-channel interpolation, re-referencing, and ICA. It reviews standard analysis approaches, including ERPs, time-frequency analysis, connectivity, and source localization. It also examines machine-learning methods, with attention to cross-subject generalization, data leakage, evaluation, and reproducibility through BIDS-EEG and FAIR data practices.
Parsa Razmara, Woojae Jeong, Aditya Kommineni et al.
In briefAct First, Reason Later: Faster On-Policy Distillation for Multi-Turn Agents
SummaryThis paper addresses the slow training of on-policy distillation (OPD) for multi-turn language agents, where lengthy reasoning before each action delays environment interaction. The proposed ActFirst-OPD framework decouples acting from full response generation, using reference-conditioned inverse dynamics to execute actions and switching to autonomous prediction when trajectories deviate from the reference. Experiments with 0.6B, 1.7B, and 4B Qwen3 students report average wall-clock speedups of 2.3x on ALFWorld, 1.8x on WebShop, and 4.9x on ScienceWorld over Vanilla OPD.
In briefPixel-wise Exposure for More Robust In-Vehicle Remote Heart-Rate Sensing
SummaryRemote photoplethysmography (rPPG) measures heart rate from a camera, but uniform exposure in bright car cabins overexposes or underexposes parts of the face. The paper proposes PixExpo, which captures frames under a cyclic exposure schedule and fuses them per pixel by choosing the value closest to an rPPG-motivated target intensity, without sensor changes. On the new MEX-Drive dataset of 48 participants in real driving, it reduces mean absolute error from 13.94 to 6.73 bpm and raises success rate from 25.95% to 63.24% versus manufacturer-default auto-exposure.
In briefSAKI: Teacher-Supervised On-Policy Distillation with Maximal-Coupling Routing
SummaryThis paper proposes SAKI, a method for on-policy distillation in which a student learns from its own generated trajectories guided by a teacher. It combines a KL-constrained teacher-guided rollout with maximal coupling and uses accept and correction events to assign token-level supervision. Accepted positions keep sampled-token reverse-KL supervision, while correction positions receive direct supervision on the teacher's most probable token. The authors report improved Mean@8 and Pass@8 over a matched teacher-guided baseline on seven math reasoning benchmarks for both 1.7B and 0.6B students.
In briefSecond-moment stochastic approximation: optimization methods using both mean and variance estimates
SummaryClassical stochastic approximation methods rely only on first-moment (mean) estimates of a random regression function. This paper studies methods that use estimators of both first and second moments, which include modern optimizers such as Adam and Muon. The authors derive these methods through optimal preconditioning for solving matrix equations and propose a two-stage convergence analysis. They show that practical methods, using estimated moments, converge almost surely to a neighborhood of the target solution, with size depending on estimator bias and variance. Concrete bounds are derived for Muon and a spectral variant of Adam.
In briefScaling video generation for reasoning: how costly is it?
SummaryThe paper asks whether scaling video generation models lets them infer hidden information from past frames, and at what compute cost. Using a controlled 2x2x2 Rubik's Cube benchmark with exact simulator ground truth, models predict visible sticker configurations after nine prescribed moves. Validation MSE follows approximate power-law scaling, but lower MSE does not reliably indicate reasoning ability, and smaller models reach higher state accuracy with limited compute.
In briefBeyond Legibility: Benchmarking Text Rendering and Editing in Video Generation
SummaryThis paper introduces VidScribe, a benchmark for testing how well video generation models render and edit visual text. It covers four tasks: writing text from language, transferring text from a reference, keeping text stable during motion, and editing text locally, using 803 human-verified samples. Tests on 11 commercial and open-source systems show that editing is the hardest task, and that the benchmark's preference optimization measurably improves text generation.
In briefSimpleARM: a training-free memory layer that lets frozen generalist robot policies keep task-relevant state from past interaction
SummaryThe paper introduces SimpleARM, a training-free memory layer that attaches to frozen generalist robot policies. It uses the task instruction to decide what to monitor, maintains compact typed state online with frozen perception tools, and retrieves that state when a subgoal depends on history. On RoboMME, a benchmark of 16 memory-dependent manipulation tasks across three policy seeds, SimpleARM reaches 67.17% mean success versus 44.51% for the strongest non-oracle baseline.
In briefText2Sim: Agentic Physics Simulation Generation Using Distilled Expertise
SummaryThis paper presents Text2Sim, an agentic pipeline that turns a text-only request into an executable, editable physics simulation built on Genesis. It combines a Planner, specialized Writers, asset-generation tools, and an independent Critic, using Debug Cards distilled from graphics demonstrations to guide repair. On 42 held-out prompts, Text2Sim scores higher than four state-of-the-art baselines on automatic physical and visual metrics, and significantly more participants preferred it in blinded user studies.
Xiaoyu Xiong, Tsun-Hsuan Wang, Yi-Ling Qiao et al.
In briefSEED: Faster LLM Text Generation by Reusing Internal Representations
SummaryThis paper proposes SEED, a method that speeds up large language model inference by drafting tokens from the model itself. It treats a standard decoder-only transformer as an implicit encoder-decoder, reusing deep representations computed during verification to draft tokens cheaply. Reported experiments on 4B-scale models show up to 2.7x average speedup, running 28% faster than EAGLE-3, with generation quality preserved or improved.
In briefReinforcement fine-tuning for cooperative multi-robot vision-language-action models
SummaryThis paper studies reinforcement learning for fine-tuning vision-language-action models so that multiple robots can cooperate. The authors propose a three-stage pipeline: initialization-aware data collection that uses human demonstrations only when the pretrained model repeatedly fails, offline fine-tuning on per-agent trajectories with credit filtering, and online RL in the frozen model's latent noise space. The method is evaluated on RoboTwin, RoboFactory, and real-world tasks with two Franka robots, reporting average success-rate gains of 23.1%, 16.4%, and 44% respectively.
WhyCombines credit-filtered tuning and latent-space RL, validated on 11 tasks with substantial gains over prior multi-robot VLA baselines.
Ruixiao Xu, Wong Lik Hang Kenny, Zhiqian Liu et al.
In briefReinforcement learning can make models ignore the requested answer format
SummaryThis paper studies whether reinforcing one answer-reporting convention in RLVR training reduces a model's adherence to a different requested format. On GSM8K, Qwen2.5-7B trained with boxed-format rewards dropped sharply in hash-format compliance in four of five seeds. The authors argue that convention-matched accuracy alone does not capture post-training behavior, and that current-request adherence should be evaluated alongside it.
In briefTransformers stop reasoning too early; a tiny LoRA extends it
SummaryThe paper finds that pretrained transformers use little depth to follow in-context references, reliably tracing only about 1.4 to 3.6 lines across thirteen base models. With all weights frozen, a rank-8 LoRA at one early layer extends this computation. For example, Qwen3-8B's exact accuracy on 24-line chains rises from 15.5% to 99%.
WhyShows pretrained transformers underuse depth; a frozen-model rank-8 LoRA lifts 24-line chain accuracy from 15.5% to 99%, with solid multi-model experiments and notable subfield impact.
In briefMaking Analog Training Scale: Co-Designing Mapping, Optimizer, and Converters
SummaryAnalog in-memory computing performs matrix operations where weights are stored, offering a route to model training, but hardware non-idealities make scaling to modern deep models hard. This paper presents a system-algorithm co-design that optimizes weight mapping, pairs a preconditioned optimizer with threshold-triggered open-loop pulsing, and aligns converter ranges to reduce quantization errors. Using architectural simulations calibrated with electrochemical RAM measurements, the framework scales Transformer training to 123M parameters, with validation loss scaling as N^-0.231, comparable to N^-0.238 for digital training.
Zhaoxian Wu, Tayfun Gokmen, Omobayode Fagbohungbe et al.
In briefMemEvo: Automatically discovering memory methods for streaming video understanding
SummaryThis paper addresses streaming video understanding, where vision-language models must compress an endless visual stream into bounded memory before knowing future queries. The authors treat memory design as a search over executable memory programs, defined in a lightweight domain-specific language. They introduce MemEvo, an LLM-driven framework that iteratively generates and refines candidate programs using experimental feedback. Experiments on StreamingBench and OVO-Bench report strong performance and substantial context and inference efficiency.
In briefSafeCoEvo: Test-Time Co-Evolution of Safety Harnesses and Guards for LLM Agents
SummaryThis paper proposes SafeCoEvo, a test-time framework that lets an external safety system for LLM agents keep adapting using experience gathered during deployment. It works on two timescales: S-Harness quickly turns recent runtime experience into updatable explicit safety knowledge, while GuardVPO gradually consolidates accumulated experience into parametric risk-judgment ability. According to the paper, compared with the strongest baseline, it reduces the unsafe outcome rate by 10.05% and improves the task success rate by 12.15%.
In briefFactorized Scheduling Principle: interpretable, transferable scheduling rules learned via additive functions
SummaryThis paper proposes a factorized scheduling principle (FSP) framework for learning interpretable and transferable scheduling rules. It represents system states as condition distributions and decomposes a global scheduling principle into additive univariate and pairwise components with identifiability constraints. Experiments on synthetic and realistic scheduling tasks show strong performance, interpretability, and zero-shot generalization across different system scales.
In briefDriftOPD: Making One-Step Robot Control Policies Learn Long-Term Task Success
SummaryThe paper introduces DriftOPD, a framework that distills long-horizon task behavior into one-step vision-language-action (VLA) robot action experts without a teacher model or policy rollouts. It splits the sequence-level reverse KL divergence into a chunk-level term and a future-potential term, optimized with a one-step drifting objective and a Q-function critic learned from offline demonstrations. Experiments across several VLA architectures in simulation and real-world manipulation show it generally outperforms one-step distillation baselines and approaches multi-step teacher performance.
In briefMemory-guided audio enhancement to help audio-language models hear a target sound amid noise
SummaryAudio large language models can struggle to understand a target sound when background noise and other sources are mixed in. This paper proposes LTM-AE, which stores clean reference audio representations per sound category as long-term memory and blends reconstructed audio tokens with the originals, without training the model. Tested on twenty sound categories and three models with three interfering sources, the method improves accuracy for the specified target over raw mixtures.
In briefDivide and Inject: Can Agents Reconstruct Attacks from Fragments?
SummaryThis paper introduces AdaLCPI, an adaptive long-context prompt injection method that splits an attack objective into incomplete fragments embedded in content an agent retrieves through its tools. A reconstruction cue asks the agent to combine them, and the fragments and cue are iteratively refined with OpenEvolve. The authors report a 61.4% macro-average attack success rate, versus 32.8% and 30.0% for the compared baselines. They argue that safety evaluations should test whether agents remain robust when harmful objectives must be reconstructed from incomplete fragments.
In briefVisual sensitivity versus claim retractability in multimodal RL credit assignment
SummaryThis paper studies how reinforcement learning with verifiable rewards assigns credit to visual claims in vision-language models. Before training, 27.81% of Qwen2.5-VL-7B's correct answers on four multimodal benchmarks contained at least one visual claim the image does not support, and outcome-level rewards credit these claims too. The authors propose Persistence-Aware Credit Gating (PACG), which attenuates positive credit for unusually persistent visual claims without requiring labels or extra inference cost, and report gains on nine benchmarks.
In briefCheckpoint Selection Gains in Supervised Fine-Tuning: Separating Three Claims
SummaryThis paper studies how to choose one checkpoint to keep from several produced during supervised fine-tuning (SFT). The authors treat selection as a finite-information decision and separately test whether more validation data help, whether selection rules beat negative log-likelihood selection, and whether they beat keeping the final checkpoint. Experiments on mathematical and Commonsense trajectories show gains from larger validation budgets and over NLL selection, while gains over the final checkpoint remain unresolved.
In briefConvergence and sampling trade-offs for diffusion models on curved spaces with nonnegative curvature
SummaryThis paper studies how quickly diffusion generative models on Riemannian manifolds converge and how costly sampling is. The authors propose a framework that separates score discretization from Brownian-motion simulation, allowing several geodesic random-walk steps between score evaluations. Under nonnegative Ricci curvature and an exact Brownian-motion simulation oracle, they show that about O(d/ε²) score evaluations achieve KL divergence ε², matching the rate known for Euclidean diffusion models.
In briefAffectReveal: Recognizing Emotions Through the Events Behind Visible Reactions
SummaryThis paper argues that visible reactions such as tears can signal different emotions, so images alone are often insufficient for emotion recognition. The authors define Event-Grounded Emotion Recognition and build EGER-Bench, with 10,052 videos and 10,734 images across 11 emotions. They propose AffectReveal, a tuning-free framework that recovers the affect-determining event and cross-checks it against face-masked evidence, improving average UAR across three models and four input settings.
In briefThinkingGuard: a safety model that detects hidden multimodal risks through step-by-step reasoning
SummaryMultimodal large language models can produce unsafe outputs when harmless text and neutral images combine to form a risk, which the authors call implicit risk. They built TriggerBench, a 5,600-instance dataset using counterfactual contrastive pairs to model such risk combinations. They then trained ThinkingGuard, a guard model, with a step-supervised structured reasoning framework, and report strong results on standard and implicit safety benchmarks.
In briefFM-ReID: Selective Competitive Token Routing for Object Re-Identification
SummaryThis paper addresses object re-identification, where different identities can look alike and their distinguishing cues are localized. It proposes FM-ReID, an end-to-end framework in which multiple mining queries and a residual query compete for dense DINOv3 tokens, with above-prior selection retaining the tokens each query is allocated. The multi-query descriptors are trained jointly with a holistic representation, without fixed spatial partitions. The method reports strong results on animal, person, and vehicle ReID benchmarks.
In briefHiTS-CL: A continual learning method for predicting future facts in evolving knowledge graphs over long time spans
SummaryExisting temporal knowledge graph reasoning methods train once on an early part of the timeline and then use a frozen model for future timestamps, which degrades over long horizons as the data changes. This paper frames extrapolative reasoning as continual learning over streaming snapshots and proposes HiTS-CL, a backbone-agnostic framework that uses continual fine-tuning for current dynamics, multi-teacher adaptive distillation for stable knowledge, and a selective memory of recent and frequent facts. HiTS-CL was integrated into five representative backbones and tested on four benchmark datasets. The authors report consistent gains in extrapolation accuracy and reduced long-horizon degradation compared with continual-learning baselines.
WhyReframes temporal KG extrapolation as continual learning, combining fine-tuning, multi-teacher distillation and memory; validated on five backbones and four datasets.
In briefWhy medical vision-language models underuse their vision encoders, examined through dermatology
SummaryThis paper studies why medical vision-language models (VLMs) perform worse than their own vision encoders in dermatology, where the MedSigLIP encoder beats MedGemma by an average of 10.26 percentage points. The authors test three hypotheses and analyze internal attention, finding that a describe-then-decide prompt raises vision attention by 30-40% during generation. They propose label-free prompting combined with low-label encoder-assisted reranking, keeping the VLM frozen, and validate it across five VLM backbones.
In briefMAADBench: A Refreshable Benchmark for Anomaly Detection in Multi-Agent Systems
SummaryThis paper introduces MAADBench, the first refreshable benchmark for detecting anomalies in multi-agent systems built on large language models. It addresses task leakage and expiring traces by combining sampled generative tasks, trace generation under configurable LLM backbones, and automatically provided step-level labels. The authors release MAADBench-Full, a dataset of 5,200 step-labeled traces across five LLM backbones. Evaluating 25 anomaly detection methods shows that current approaches depend heavily on supervision, struggle with subtle anomalies, and lack robustness across backbones.
WhyProposes a refreshable multi-agent anomaly detection benchmark, systematically evaluating 25 methods across five LLM backbones and revealing clear limitations.
In briefSERA: A Rollout Allocation Method That Equalizes Gradient Scaling in Maximum Likelihood Reinforcement Learning
SummaryThis paper addresses how finite rollout budgets in Maximum Likelihood Reinforcement Learning (MaxRL) shrink each prompt's likelihood gradient in a way that depends on its success rate. The authors propose SERA, which redistributes a fixed rollout budget so that these finite-rollout scaling factors are roughly equal across prompts. In a controlled ImageNet setting, SERA aligns more closely with exact likelihood gradients, and on maze navigation and mathematical reasoning it improves multi-sample solution coverage over MaxRL under matched training budgets.
In briefGrounded Revision or Prior Injection: Testing Retrieval-Augmented Patent Claim Amendment
SummaryThe paper tests whether retrieval-augmented generation grounds patent claim revisions or merely injects templates, using 7,385 USPTO prosecution cases with examiner rejections and cited prior art as ground truth. It releases a corpus, a seven-probe test battery, and a deterministic five-channel metric that needs no LLM judge. Across 9,600 pre-registered calls on four LLMs, no model showed detectable classical prior-injection behavior, and retrieval effects were small and direction-inconsistent.
Josepha Michiko Leo, Hyun-seok Min, Yehoon Jang et al.
In briefInteractive-Policy Distillation (IPD): A Student-Teacher Method That Alternates Proposing and Verifying
SummaryThis paper studies how a student model can learn from a teacher's token-level feedback on its own reasoning trajectories. The authors note that standard on-policy distillation (OPD) can suffer when the student drifts far from the teacher, making supervision unreliable. They propose Interactive-Policy Distillation (IPD), in which student and teacher alternately act as proposer and verifier to produce mixed-source trajectories, with supervision applied according to each token's source. On math reasoning tasks, IPD-trained students outperform OPD while using about a quarter of the training examples and steps.
In briefSCCM: A Coarse Matching Method for Dense Feature Correspondence on Equirectangular 360° Panoramas
SummaryThis paper addresses dense feature matching between 360° panoramas stored in equirectangular projection, which introduces seam, latitude-dependent stretch, and non-uniform pixel area distortions. The authors show that perspective-trained matchers degrade on this format and propose SCCM, which adds spherical positional attention and area-aware covisibility priors to the coarse matching stage. On Matterport3D, SCCM raises PCK@1° from 0.230 to 0.275 and outperforms ERP-native EDM and an ERP-retrained RoMa V1 under a unified protocol.
In briefDraftTrace: A Multi-View Analytics Environment for Writing with Integrated AI
SummaryThis paper introduces DraftTrace, a writing environment that records the final product, the writing process, and interactions with an integrated AI assistant, reconstructing how a document develops over time. It was deployed in a graduate NLP course with 81 students and compared sessions against LLM-generated responses entered by automated tools and against copy-typed responses. Product measures distinguish differences in text formulation, process measures distinguish how text was entered, and combining both helps identify cases such as copy-typing. Students used the assistant to clarify questions early in writing and to verify answers later.
In briefReactive real-time robot control via asynchronous action-distribution alignment
SummaryThe paper examines whether asynchronously executing a vision-language-action (VLA) robot policy yields the same action distribution as the original model. It finds that for non-Markovian demonstrations, asynchronous execution can produce a fundamentally different distribution, limiting reactivity. The authors propose Recursive Flow-Field Distillation and Propose-Resolve to align the asynchronous policy with the original VLA. They report matching the original VLA on LIBERO and retaining about 80% of its success on RoboMimic, roughly 30 points above existing asynchronous methods.
Moritz Zoellner, Reece O'Mahoney, Ioannis Havoutis et al.
In briefWhen Updating Stops Being Learning: Rethinking LLM Self-Evolution via Learnable Information Gain
SummarySelf-evolving large language models improve using their own generated data but often degrade after a plateau. The paper proposes a learnable information gain measure, estimated by fitting a small language model to the previous round and scoring new data by negative log-likelihood. Building on it, the authors introduce ATRI, which reweights samples within a round and halts training across rounds when information gain stays low. The authors report experiments on popular datasets demonstrating its superiority.
In briefHow Retrieval Models Respond to Political and Dialect Signals in Queries
SummaryThis paper examines whether dense retrieval models are biased by identity signals in user queries, namely political ideology and dialect, in political news and consumer-health domains. Across five dense retrievers and a sparse baseline, all systems retrieved articles matching the query's political lean and performed worse on African American Language than on White Mainstream English. Probing analyses indicate these gaps reflect identity signals beyond surface vocabulary, recoverable from query embeddings.
WhySystematically shows dense retrievers are biased by user political lean and dialect, validated on synthetic and naturalistic queries beyond lexical explanations.
Andrew Tang, Nicholas Deas, Kathleen McKeown et al.
In briefHierarchical Utility Calibration for Structured Multiclass Decisions
SummaryThis paper studies how Utility Calibration (UC) for multiclass probabilistic prediction behaves when labels follow a hierarchy. The authors show that the gap between realized and predicted mean utility decomposes exactly into contributions from internal nodes of the label tree, and that these contributions can cancel, so small overall UC can hide large utility errors in parts of the hierarchy. They propose Hierarchical Utility Calibration (HUC), which evaluates each node's contribution before summing, and HUC-Boost, which updates only violated internal nodes; both come with theoretical guarantees and finite-sample evaluation.
In briefPredicting what happens after a fall: testing video world models on physical events
SummaryThis paper tests whether video world models can anticipate what happens next when a release or impact has just occurred but its outcome is hidden. The evaluation uses 62 real-world free-fall recordings and 124 clips across three object types, annotated with release and impact times and ground-truth trajectories. Some models produce the events at high rates but often start them late, while others usually keep the pre-event state with little measurable consequence. A 15-participant human study is included for comparison.
In briefTriadic Linear Attention: A Three-Dimensional Memory for Long Sequences
SummaryRecurrent neural networks store past context in a fixed-size memory, and linear attention enlarges this memory into a matrix. This paper proposes triadic linear attention, which writes the three-way outer product of two keys and a value into a third-order tensor state and reads it by contracting both keys with two queries. When applied to Gated DeltaNet and scalar-gated linear attention, it substantially improves long-context language modeling and recall, outperforming alternatives that simply enlarge the state.
Oliver Sieberling, Bharat Runwal, David Jin et al.
In briefSCOPE: A method that predicts full physical fields from sparse measurements
SummaryThis paper proposes SCOPE, a method for recovering complete physical fields governed by PDEs from sparse observations. It combines full-field latent prediction with physical reconstruction, using a shared decoder trained on both. Across five PDE settings, the authors report that SCOPE outperforms mask-aware neural operators on all ten forward and inverse tasks and achieves lower errors than diffusion-based solvers such as DiffusionPDE and FunDPS.
In briefImproving Context Compression for Long-Running Agents via Counterfactual Continuations
SummaryLong-horizon agents must compress growing interaction histories, but compression quality is ultimately judged by downstream execution. The paper introduces PAIR, which uses matched counterfactual continuations with and without compression from the same agent state to find individual compressions that degrade later execution. PAIR then revises relevant sections of a fixed structured compression template. The authors report that PAIR achieves the strongest cross-run reliability among compressed methods on main benchmarks without modifying the downstream agent.
In briefTesting whether medical AI models behave consistently when deployed across distributed systems
SummaryThe paper notes that medical models evaluated in centralized settings may produce different outputs when served with distributed inference techniques such as tensor parallelism and mixed precision. The authors propose a testing framework and a benchmark sensitive to distributed execution that compares a centralized reference with matched distributed deployments. Experiments on language, vision, and multimodal medical models show measurable output disagreements, with test success rates of 0.21 to 0.43 for single-modality visual settings and 0.32 to 0.98 for multimodal models.
In briefPDE-OBS: A controlled benchmark for how measurement patterns affect physical-field models
SummaryThe paper presents PDE-OBS, a benchmarking platform for reconstructing and forecasting physical fields under different observation patterns. It includes 560,000 fields and trajectories from seven PDE families and seven baseline methods. The authors find that error is higher when test observation patterns differ from matched patterns across all 49 PDE-method pairs, and that mixed-pattern training reduces transfer errors but usually does not match destination-trained references.
In briefDistilling a privileged safety barrier teacher into an RGB-only filter for dynamic navigation
SummaryThe paper proposes a teacher-student visual distillation framework that transfers the safety behavior of a privileged Control Barrier Function teacher, which uses ground-truth states in a dynamic Gaussian Splatting environment, to an RGB-only student filter. The student takes a short RGB history, robot velocity, and a nominal action to output a safe action, requiring no explicit 3D reconstruction or online rendering at deployment. Experiments reportedly show it outperforms visual CBF baselines and improves safety of RGB-based navigation under dynamic obstacle motion.
In briefLIBERO-MAX: Can Robot Policies Adapt When the World Changes Mid-Task?
SummaryThe paper introduces LIBERO-MAX, a benchmark of 8,000 paired cases covering eight types of changes to geometry, observations, appearance, clutter, and paths. Each pair compares task execution with and without a mid-task event, keeping everything else fixed so that event-caused failures can be separated from pre-existing ones. Across fourteen current policies, events reduced success by 11.0 to 25.7 percentage points, with geometry and observation changes showing shared vulnerabilities.
In briefLarge-scale factor analysis finds machine intelligence is only partly interpretable
SummaryThis paper applies factor analysis, a psychometric method, to 13,251 evaluation scores from 1,618 language models across 456 text-only benchmarks to test whether a general intelligence factor exists. At the most generous estimate, a general factor accounts for 70.8% of performance variance, and much less in most solutions. Content-similar benchmarks do not necessarily cluster, and the factor is not dominated by a common theme or well proxied by standard intelligence benchmarks. The authors conclude this does not support treating general intelligence as a tangible target for language model development.
WhyFirst large-scale factor analysis of 1,618 language models across 456 benchmarks challenges the general-intelligence assumption, with robustness checks across imputation methods.
In briefBRIDGE: A bilevel method that jointly trains a search retriever and an LLM agent with retrieval-aware credit assignment.
SummaryThis paper studies agentic reinforcement learning, where large language models interleave search and reasoning, and notes that existing methods treat retrieved evidence as environment input, misattributing retrieval failures to the LLM. The authors find that adapting the retriever before the policy yields larger reward gains, and formulate joint training as a bilevel optimization problem. They propose BRIDGE, a memory-efficient first-order method, which the authors report achieves the highest average accuracy across seven open-domain QA benchmarks with 3B and 7B backbones and performs well on medical QA benchmarks.
In briefPhysics-guided fully convolutional model for fast 3D microstructure evolution prediction
SummarySimulating how three-dimensional microstructures evolve with phase-field methods is computationally expensive. This paper proposes an efficient, scalable framework that directly predicts complete 3D microstructure sequences using fully convolutional spatiotemporal learning. A discrete Cahn–Hilliard residual is added during training to guide the model toward the governing dynamics, without changing inference. On 3D spinodal-decomposition data, average 3D structural similarity stays above 0.97 over the nominal horizon, with more than 30-fold speedup over the reference spectral phase-field solver.
In briefAVIO: Learning to Add and Remove Sounding Objects in Audiovisual Scenes
SummaryThis paper introduces AVIOBench, a dataset of 37.9 hours of paired audiovisual examples covering 1,878 target-object names, linking each object's visual presence and sound through a shared identity and visual mask. Building on it, the authors propose AVIO, which adapts a pretrained text-to-audiovisual generation model via source-conditioned feature modulation to learn both adding and removing sounding objects. The paper reports that quantitative and qualitative evaluations show effective audiovisual object removal and addition, with optional reference guidance providing appearance and placement control for addition.
In briefUsing AI-generated multimodal tutoring transcripts to move past the learning system wall
SummaryPast research using log data has struggled to generalize models of student learning across platforms, a problem the authors call the "learning system wall." They introduce an AI-driven system that turns screen-recording videos of remote tutoring into unified screenplay-style transcripts with audio dialogue and annotated learning actions. The paper outlines a planned method for aligning these transcripts with MATHia learning logs and classifying student learning processes, and discusses challenges and possible solutions.
Danielle R. Thomas, Marie Cynthia Abijuru Kamikazi, Ashish Gurung et al.
In briefInterBias-SV: A Benchmark for Compound Conditions in Speaker Verification
SummaryThis paper examines how speaker verification systems perform when noise, channel distortion, and speech variation occur together, and whether their individual effects add up. It frames the question as a four-term comparison of joint error, two marginal errors, and a common reference, using 4,068 scored records across 17 experiments and six speech corpora. The reported mean contrasts are descriptive and do not establish that the effects are additive. The authors also discuss interpretation issues, including near-chance EER and unstable ratios of demographic gaps.
Kamel Kamel, Hridoy Sankar Dutta, Keshav Sood et al.
In briefReimagine Video Dynamics: a framework that separates and edits how a video moves from what it looks like
SummaryThis paper introduces Reimagine Video Dynamics (RVD), which learns a compact, editable dynamics token through self-supervised reconstruction, capturing how a scene evolves rather than its appearance. Because the token is disentangled from visual context, video motion can be edited directly while the visuals are preserved. Experiments show effective dynamics editing, training-free retiming, and appearance-controlled re-rendering.
In briefBenchmarking vision-language models on checking synapses and neurons in connectomics images
SummaryThis paper tests vision-language models (VLMs) on the decisions annotators make when reviewing electron microscopy images, such as whether a synapse is present and its polarity, or whether neurons were wrongly split or merged. It evaluates 19 open and 2 closed models for synapses and 3 open and 2 closed models for neurons. Most models performed at chance without examples, while LoRA fine-tuning on a few thousand labels brought open models level with specialist models. On unseen species, the best adapted VLMs outperformed specialist models at identifying merge errors.
In briefLLMs Learn to Evade Latent Monitors from Prior Feedback Alone: language models learning to hide internal activity from detectors using feedback.
SummaryThe paper studies whether latent-space monitors, which inspect LLM internal activations to detect undesired behavior, can be evaded by agents. Models inferred the monitor's decision rule from prior verdicts and edited activations accordingly. Scaling edits by 8x cut the monitor's true positive rate from 100% to 27%, and a rank-1 LoRA reduced it to 4% on held-out concept monitors while retaining benchmark capabilities.
WhyShows models learn to evade latent monitors from feedback alone, with scaling and LoRA experiments supporting the claim and a significant AI-safety impact.
Hugo Lyons Keenan, Christopher Leckie, Sarah Erfani
In briefAdaptArena: A Benchmark for Web Agents' Test-Time Personalization
SummaryThis paper introduces AdaptArena, a benchmark of 480 tasks that tests whether web agents can infer a user's implicit preferences from past behavior. It also presents AdaptiveAgent, a retrieval-based framework for standardized evaluation. Agents with ground-truth preferences reach an 82.92% success rate, while the evaluated LLM agents reach at most 15.62%. The authors also find that correct preference inference alone does not guarantee success, since execution failures in web interactions remain a major bottleneck.
In briefOptimal sample complexity for learning policies across multiple rewards
SummaryThis paper studies a finite-horizon Markov decision process with unknown transitions and several known reward functions, aiming to learn an ε-optimal policy for each reward using only online episodic interaction. The authors design a provably efficient algorithm with a minimax sample complexity of O(SAH³/ε² · log M) up to polylogarithmic factors, with no extra burn-in cost. The method combines optimistic value estimates adapted from MVP to reward switching, conservative policy evaluation with fresh replay samples, and gap-based multiplicative weights that allocate learning across rewards.
WhyAchieves a minimax-optimal sample complexity bound for multi-reward MDPs matching lower bounds, a notable theoretical advance within the subfield.
In briefExtrapolating RL-induced representation shifts in on-policy distillation
SummaryThe paper proposes RIDE, a method for on-policy distillation that extrapolates the change induced by reinforcement learning directly in representation space. At each layer and token position, it computes the residual between the teacher and its pre-RL checkpoint and pulls the student's hidden states toward targets displaced beyond the teacher along this residual. Across four base/RL-teacher pairs with different scales, architectures, and pre-training lineages, RIDE approaches or exceeds the RL-trained teacher and outperforms output-space extrapolation.
In briefDSB-DG: Benchmark for Document Grounding and Hallucination in Voice Agents
SummaryThis paper introduces DSB-DG, a benchmark that tests whether voice agents give answers faithful to external documents across five professional domains. It contains 1,636 adversarially verified question-answer pairs drawn from 50 documents and supports automatic scoring of grounding accuracy, hallucination, and latency. The cascaded ASR-LLM-TTS pipeline scored highest on grounding accuracy, with Gemini-Live and GPT-Realtime close behind, while open-weight systems showed context collapse and multi-turn decay. The authors find that grounding degrades with longer context and conversation load, and that failures often appear as unsupported generations rather than abstentions.
In briefHuman-AI Collaboration: From Paradoxes to Patterns (a paper on how people and AI work together)
SummaryThe paper notes that humans and AI systems tend to perform better when collaborating than when working alone. It examines autonomy and initiative as two design dimensions and the collaboration patterns they generate, analyzing the internal tensions within underlying problems through a paradox perspective. It describes a process for surfacing these tensions and derives four collaboration patterns: Instruction, Delegation, Assistance, and Co-creation.
In briefLearning to Harvest Without Collapse: A Lagrangian Framework for Regenerative Commons
SummaryThis paper frames the tragedy of the commons as a constrained multi-agent Markov game in which a designer sets a depletion budget that explicitly requires resource preservation. It proposes a nonstationary Lagrangian framework that builds policy sequences from solutions to unconstrained games or cooperative control problems, with feasibility and approximate optimality guarantees. Experiments with constrained IPPO and MAPPO in a Gordon-Schaefer fishery examine how depletion budgets affect stock retention, harvest rewards, and price adaptation.
In briefSafety guard models are overconfident under attack where their base models are uncertain
SummaryThis paper evaluates five guard models used to classify prompts for safety. Although several are nearly calibrated on clean inputs, adversarial attacks degrade their calibration by an order of magnitude, making false negatives appear as high-confidence errors. Comparing each guard with its base language model, the base model typically shows uncertainty on the same inputs where the guard fails, and layer-wise analysis localizes this divergence to later layers.
In briefSimilar Choices, Different Attention: How Humans and Vision-Language Models Link Sounds to Shapes
SummaryThis paper compares humans and vision-language models on cross-modal associations, such as pairing a pseudo-word with round or sharp shapes. Both were given identical stimuli, and the researchers recorded human choices and eye movements from 53 participants. Some larger models matched human choices, but their attention matched human gaze less closely than a fixed center-bias baseline. The authors conclude that matching human choices or gaze patterns alone is not sufficient evidence of human-aligned processing.
In briefGoing Beyond State-Reaching: Learning Abstractions for Intrinsically Motivated Option Discovery
SummaryThis paper proposes an option discovery algorithm that, instead of having each subgoal cover all aspects of the state, selects a small relevant subset of features for each subgoal. The resulting options generalize more broadly and speed up exploration in sparse-reward environments. The method is evaluated on three image-based, sparse-reward domains, including the Atari game Montezuma's Revenge.
Akhil Bagaria, Anita De Mello Koch, George Konidaris
In briefDisCoMBO: Distributional-conformance method for expert-guided black-box optimization
SummaryThis paper addresses how to incorporate expert knowledge into sequential model-based optimization using Probabilistic Circuits as surrogates. It introduces the Distributional Conformance Score (DisCo), a bounded measure of model surprise that also assesses how well external knowledge conforms to model evidence. The authors present DisCoMBO, prove it is a zero-regret algorithm, and demonstrate it on AutoML, materials, and wind park optimization benchmarks.
WhyProposes a distributional conformance score on probabilistic circuits unifying uncertainty and expert knowledge injection, proves zero-regret, and validates across several benchmarks.
In briefStaircase Policy: streaming inference for robot world-action models with large action chunks
SummaryThis paper introduces Staircase Policy (S-WAM), a streaming inference and training framework that converts a flow-matching vision-language-action model into a JEPA-style world-action model. It splits a large action chunk into sub-chunks at staggered denoising stages, so near-term actions run as soon as they are ready. At each sub-chunk boundary, the future latent is re-predicted from the latest observation to update unexecuted actions. The authors report 97.7% on LIBERO, 87.9% on LIBERO-Plus, and 3.62 times the throughput of conventional execution at comparable accuracy.
In briefRethinking reasoning paths as phase-structured trajectories: a method for comparing a model's solution attempts within each question
SummaryLarge language models often solve problems by generating multi-step reasoning paths, but how to analyze their hidden states along these paths is unclear. The authors propose PAIR, which samples several trajectories per question, aligns variable-length paths into shared relative phases, and compares successful and unsuccessful trajectories only within the same question and phase. They report that standard across-question correctness probes lose much of their predictive power under within-question evaluation, while PAIR improves within-question trajectory ranking and Best-of-N selection. Phase-wise steering experiments further show the learned directions can change generation outcomes.
In briefLearned chord embeddings reveal tonal structure linked to musical tension
SummaryThis paper studies whether skip-gram chord embeddings learned from chord sequences capture the tonal structure of Western tonal music. The authors compare embeddings trained with and without transposition-based augmentation using geometric, functional, and tension-related evaluations. Augmented embeddings show strong transposition equivariance and recover a clear circle-of-fifths structure. Embedding-based tension measures correspond to matched tonal measures and moderately align with human tension profiles.
Maral Ebrahimzadeh, Gilberto Bernardes, Sebastian Stober
In briefFisher-IRG: a Fisher-based method for local invariant representation geometry in language and vision models
SummaryThe paper proposes Fisher-IRG, which measures local representation directions by their predictive sensitivity rather than Euclidean distance. Across language and vision models, it shows stronger selectivity between semantic and nuisance variation and generally more reproducible subspaces than covariance-based geometry. Semantic effects localize to the Fisher-derived subspace, and held-out separation and retrieval results indicate the geometry generalizes beyond the discovery neighborhoods.
In briefMemory Consolidation Flattens Time Cues in Stored User Facts
SummaryThe paper shows that long-term memory systems often turn progressive statements such as "I am driving a Peugeot" into simple-present notes, losing the cue that the fact is ongoing. The authors introduce the LAPSE benchmark and find that writers flatten progressive forms selectively, in 244 of 381 pairs, and never the reverse, across 11 configurations and three installed pipelines. Changing only the stored verb also shifted later readers' judgments about whether a fact still holds.
In briefPhase transition in how models store more features than their width
SummaryThis paper studies superposition, where a model represents more features than its width by letting representation vectors overlap, using Anthropic's toy model. Increasing model width drives a continuous phase transition from a partial-representation phase, where only some features are represented, to a full-representation phase where all features are represented. The authors find the critical width grows roughly linearly with the number of active features, up to a logarithmic factor, and the loss scaling changes across the transition.
In briefDynamicHOI: Physics-aware hand-object interaction reconstruction from video
SummaryThis paper reconstructs hand-object interactions from monocular RGB videos, noting that existing methods often produce visually plausible but mechanically inconsistent motion. The proposed DynamicHOI framework combines geometry-grounded diffusion refinement with coupled hand-object dynamics, transferring contact forces and modeling active hand actuation with a probabilistic prior. Experiments on three HOI datasets show improved hand and object reconstruction, and the trajectories also benefit hand world-model generation and robotic manipulation learning.
In briefChameleon: A channel-dependent state space model for multivariate time series forecasting
SummaryThis paper proposes Chameleon, a channel-dependent state space model for multivariate time series forecasting that captures fine-grained, data-dependent interactions between variables while scaling linearly with their number. It connects selective state space models with the Kalman filter, using a measurement update for cross-variable modeling, and adapts GatedDeltaNet as its backbone with additional generalization techniques. On strongly dependent ODE and PEMS datasets, it achieves the best MSE and MAE, and it outperforms baselines in most of 28 standard benchmark settings. Training-time and memory analyses on Traffic and ETT indicate competitive efficiency.
In briefReliable Parallel Decoding for Masked Diffusion Language Models: A Training-Free Speedup Method
SummaryMasked diffusion language models can predict several masked tokens at once to generate text faster, but predictions made in the same forward pass may be unreliable when committed together. The paper finds that confidence alone does not determine a safe commitment order, and proposes Reliable Parallel Decoding (RPD), a training-free method that selects candidates by layerwise prediction stability and final confidence under a cumulative entropy budget. On math reasoning and code generation benchmarks using LLaDA and Dream, RPD achieves the highest decoding throughput among the evaluated methods while maintaining or improving accuracy.
